KFUQA: Large Language Model-Driven Knowledge Fusion Framework for Zero-Shot Question-Answering

To effectively harness structured knowledge and enhance the performance of large language models (LLMs) in zero-shot question-answering tasks, this study introduces KFUQA, a knowledge fusion zero-shot question-answering framework driven by LLMs. The proposed framework consists of four key modules: Retrieval, Filtering, Textualization, and LLMs. It leverages knowledge from knowledge graphs to generate answers through two stages: relation generation and answer synthesis. Experiments conducted on the WebQSP public question-answering dataset with multiple open-source LLMs demonstrate that KFUQA significantly improves question-answering performance over methods that do no knowledge fusion or filtering, and its performance even outperforms some more complex approaches. These results underscore that KFUQA can enhance the zero-shot performance in question answering tasks of large language models without requiring model fine-tuning. Furthermore, it enables smaller-scale language models to achieve performance comparable to that of larger models.